Programação genética paralela com Pareto: uma ferramenta para modelagem via regressão simbólica

Abstract

Program induction involves the inductive discovery of a computer program that produces some desired output when presented with some particular input. An example is the symbolic regression, a modeling tool that seeks mathematical expressions of functions to fit a given multivariate data set, mapping input variables to output variables of control. The genetic programming, a subarea of evolutive computing that uses an analogy of Darwin s evolutionary theory and some ideas from the genetics field, is an automatic technique for producing a computer program widely used to solve such problems. However, implementing genetic programming is not trivial for most professionals, besides demanding high computational power. This work presents a parallel implementation of genetic programming simple to handle, optimized for computers with multicore architecture, and satisfying competitive criteria of structural simplicity model and prediction accurate model, through a special multi-objective flavor of a genetic programming, called Pareto Genetic Programing. The proposed implementation has performance gains proportional to the amount of available cores in use, and has been successfully applied to several types of regression problems.